GeoID-PINN Improves Regional Epidemic Inference with Geographic Data

Regional surveillance data for epidemics often conflates local transmission, reporting biases, initial infections, and external infection pressures, making it difficult to isolate and understand individual factors. This challenge can lead to less accurate models for disease spread. Researchers developed GeoID-PINN, a physics-informed neural network (PINN) specifically designed for Susceptible-Infectious-Recovered-Deceased (SIRD) epidemic dynamics. This model integrates spatial dependencies using a row-stochastic source-composition matrix, which assigns non-negative source weights that sum to one. This matrix is regularized using a spatial prior derived from factors like distance, adjacency, commuting patterns, or lead-lag information. In simulations, GeoID-PINN with a compatible distance prior achieved significantly lower source-composition error compared to unregularized models or those with misspecified priors, even when trajectory fits appeared similar. Retrospective evaluation on COVID-19 data from 64 Louisiana counties showed GeoID-PINN substantially reduced Mean Squared Error (MSE) and Mean Absolute Error (MAE) compared to an autoregressive negative-binomial baseline. The findings underscore that accurate trajectories alone do not guarantee correct recovery of regional dependence structures and emphasize the value of structured regularization for improved regional epidemic inference.